Fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology

Abstract

The invention relates to the field of image processing technology, and specifically discloses a fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology. The invention degrades high-quality fingerprints generated by software to obtain paired low-quality latent fingerprints and high-quality fingerprint data sets; designs a dual-branch spatial attention module and adds an enhancement network, so that the enhancement network can simultaneously learn the features of high-quality and low-quality fingerprint areas; uses reconstruction loss and multiple directional field losses and the generated fingerprint data sets to guide the training of the enhancement network, so that the enhancement network can learn the directional field features of the latent fingerprint; inputs the latent fingerprint to be enhanced into the trained enhancement network to obtain a high-quality enhancement result; and enables the enhancement network to better extract the directional field features of the latent fingerprint and the features of different quality areas, thereby improving the accuracy of the enhancement network in identifying the latent fingerprint.

Description

Technical Field

[0001] The invention relates to the technical field of image processing, and in particular to a fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology. Background Art

[0002] Latent fingerprints refer to fingerprints that remain on the surface of an object. At the crime scene, criminal investigators use specific methods to make latent fingerprints visible, use fingerprint recognition software to extract features, and compare them with the fingerprint features in the fingerprint database to determine the identity of the fingerprint and speed up the detection of the case. Since latent fingerprints are generated unconsciously, latent fingerprints often have the characteristics of small effective area, complex background, and low quality in some areas. The success rate of fingerprint recognition is much lower than that of the flat and rolled fingers commonly used. In order to improve the success rate of fingerprint recognition, it is necessary to use image processing algorithms to improve the ridge and valley contrast of latent fingerprints and repair blurred and damaged fingerprint areas, which means that latent fingerprint enhancement is needed.

[0003] The existing fingerprint enhancement algorithms can be roughly divided into two categories: one is based on filtering, and the other is based on learning. However, the existing fingerprint enhancement algorithms still have some limitations and shortcomings in practical applications.

[0004] First, the characteristics of the filtering-based method are that it is necessary to first estimate the local features of the latent fingerprint, such as the directional field and frequency field, and then design a specific filter to enhance the fingerprint. Filtering can be performed in the spatial domain, frequency domain, or both at the same time, and can filter out noise that does not conform to the estimated directional field and frequency field. However, since filtering is related to the local features at each pixel, the filter used for each pixel is different, which leads to low computational efficiency of the filtering-based method. In addition, the local features of some low-quality latent fingerprints, such as the directional field and frequency field, are difficult to estimate accurately, which further limits the latent fingerprint enhancement effect based on the filtering method.

[0005] For example, an existing Chinese patent with publication number CN106897710A discloses a fingerprint recognition system based on spectrum filtering processing technology, including: a storage module, which stores multiple fingerprint samples; a first acquisition module, which is used to acquire at least one first fingerprint; a second acquisition module, which is used to acquire at least one second fingerprint; a processing module, and an execution module connected to the processing module, which is used to perform a preset operation according to the comparison result, by calculating the Harris corner energy of each pixel of the fingerprint image, filtering out the Harris corner energy less than a predetermined threshold to obtain a binary image, dilating and corroding the binary image, respectively calculating the direction field, frequency field and curvature field of the binary image, and using a filter to filter the binary image to obtain the preprocessed fingerprint image.

[0006] Secondly, learning-based methods obtain a large amount of data through collection or software generation methods, and use this data to directly optimize the various parameters of the network. The trained network can directly enhance various latent fingerprints. However, these learning-based methods rarely take into account the quality differences in various regions of the latent fingerprint. At the same time, although the training of some networks also introduces directional field losses, these networks do not perform corresponding processing on the directional fields of the training data used, resulting in insufficient learning of the directional field features by these networks.

[0007] For example, the existing Chinese patent with the publication number CN115146670A discloses a radio frequency fingerprint recognition method and system based on data enhancement and contrastive learning, including the following steps: collecting radio frequency signals to obtain a training data set; selecting a specified number of samples as a batch training data set based on the training data set; calculating the output obtained by all samples after the convolutional neural network CNN calculation; calculating the average loss function value of each batch training process; calculating the loss function gradient, and updating the network parameters according to the gradient descent principle, repeating the above steps until the network parameter stable training ends early or the preset Epoch value iteration is completed, then the contrastive learning pre-training process based on the original data set and the enhanced data set is completed. This invention can effectively utilize the internal information of the sample and the related information between the original sample and the enhanced sample by jointly optimizing the consistency of the category samples and the consistency of the labels during the training stage, and can achieve higher individual recognition accuracy of communication radiation sources compared to the low signal-to-noise ratio radio frequency fingerprint recognition method that only relies on data enhancement. Summary of the invention

[0008] In view of the above problems, the present invention proposes a fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology. The specific technical scheme is as follows: The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology comprises the following steps: Step 1, obtaining a paired low-quality latent fingerprint set and a high-quality fingerprint set: obtaining a high-quality fingerprint image set, degrading each high-quality fingerprint once through adaptive filtering, and further degrading each high-quality fingerprint twice through Gaussian blur, morphological processing expansion and corrosion operations to generate a degraded fingerprint set of the high-quality fingerprint set, superimposing the degraded fingerprint set of the high-quality fingerprint set with the steel surface defect data set to obtain a low-quality latent fingerprint set corresponding to the high-quality fingerprint set, and constructing a paired low-quality latent fingerprint set and a high-quality fingerprint set.

[0009] Step 2: Enhanced network construction: Design a dual-branch spatial attention module and further construct an enhanced network including the dual-branch spatial attention module.

[0010] Step 3. Generate the total loss of the enhanced network: Design the reconstruction loss and each directional field loss of the enhanced network, where each directional field loss includes the directional field loss of the prediction result of the high-quality branch, the directional field loss of the prediction result of the low-quality branch, and the directional field loss of the prediction result after the corresponding features of the high-quality branch and the low-quality branch are fused, and further generate the total loss of the enhanced network.

[0011] Step 4: Enhanced network training: Use the paired low-quality latent fingerprint set and high-quality fingerprint set and the total loss of the enhanced network to train the enhanced network to obtain the trained enhanced network.

[0012] Step 5: Latent fingerprint image enhancement: Input the latent fingerprint image to be enhanced into the trained enhancement network, obtain the enhanced latent fingerprint image, and provide feedback.

[0013] Compared with the prior art, the fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology described in the present invention has the following beneficial effects: 1. In the process of degrading high-quality fingerprints generated by software to obtain paired low-quality latent fingerprints and high-quality fingerprint data sets, the present invention detects the quality of fingerprint images generated by the software, screens the generated fingerprint images, removes fingerprint images with poor quality, and retains high-quality fingerprint images, thereby improving the quality of the entire fingerprint image database, thereby improving the performance and accuracy of the enhanced network trained based on the high-quality fingerprint set and its derived low-quality latent fingerprint set.

[0014] 2. The present invention designs a dual-branch spatial attention module so that the spatial attention can fuse the different features extracted by the high-quality branch and the low-quality branch. The dual-branch spatial attention module is further added to the enhanced network so that the enhanced network can simultaneously focus on fingerprint areas of different qualities and learn the features of high-quality and low-quality fingerprint areas, thereby improving the enhancement effect of the enhanced network on latent fingerprints.

[0015] 3. The present invention guides the training of the enhancement network by using reconstruction loss and multiple directional field losses, so that the enhancement network can learn the directional field characteristics of the latent fingerprint, improves the learning ability of the enhancement network for the ridge directional characteristics of the latent fingerprint, and thus improves the accuracy of the enhancement network in identifying latent fingerprints. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0017] Figure 1It is a schematic diagram of the method flow of the present invention.

[0018] Figure 2 Schematic diagram of the dual-branch spatial attention module structure of the present invention.

[0019] Figure 3 It is a schematic diagram of the position of the directional field output and the guidance loss of the present invention.

[0020] Figure 4 It is a schematic diagram of the enhanced network structure of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] See also Figure 1 As shown, the fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology provided by the present invention includes the following steps: Step 1, paired low-quality latent fingerprint set and high-quality fingerprint set acquisition: obtain a high-quality fingerprint image set, degrade each high-quality fingerprint once through adaptive filtering, further degrade each high-quality fingerprint twice through Gaussian blur, morphological processing expansion and corrosion operations to generate a degraded fingerprint set of the high-quality fingerprint set, superimpose the degraded fingerprint set of the high-quality fingerprint set with the steel surface defect data set to obtain a low-quality latent fingerprint set corresponding to the high-quality fingerprint set, and construct a paired low-quality latent fingerprint set and high-quality fingerprint set.

[0023] As a preferred solution, the specific analysis process of step one includes: S1: setting technical parameter information for generating high-quality fingerprint images, wherein the technical parameter information includes prompt word information, generation parameter information, and model type, and inputting it into Anguli image generation software to generate a set number of reference fingerprint images.

[0024] In a specific embodiment, 5000 reference fingerprint images are generated using the image generation software Anguli.

[0025] It should be noted that the prompt words include positive prompt words and negative prompt words. Positive prompt words are used to describe the details and features of the intended high-quality fingerprint image, and may include a description of the quality, details, features, etc. of the high-quality fingerprint image. Negative prompt words are used to avoid elements that are not desired to appear in the high-quality fingerprint image.

[0026] It should be noted that the generation parameter information includes image size, resolution, style, etc.

[0027] It should be noted that there are multiple models available in the Anguli image generation software to generate high-quality fingerprint images. In a specific embodiment, the DALL-E3 model is selected to generate high-quality fingerprint images.

[0028] In another specific embodiment, the source image is input into the Anguli image generation software to generate several high-quality fingerprint images.

[0029] It should be noted that Anguli image generation software is a synthetic fingerprint generator that mainly generates fingerprint images through specific algorithms and models. It can control various parameters of fingerprint images, such as the direction, curvature, density, etc. of the fingerprint to generate different types of fingerprint images.

[0030] It should be noted that the Anguli image generation software can be replaced by other image generation software with the same function.

[0031] S2: Detect each generated reference fingerprint image to obtain the clarity and contrast of each reference fingerprint image, which are recorded as , Indicates The number of the reference fingerprint image, , by analyzing the formula Get the image quality index of each generated reference fingerprint image ,in represents a natural constant, represents the correction factor of the preset image quality index, Respectively represent the preset clarity and contrast thresholds, Respectively represent the preset clarity and contrast weights, .

[0032] It should be noted that the correction factor of the image quality index plays a role in amplifying, reducing or rounding the image quality index to facilitate calculation and analysis.

[0033] It should be noted that the definition and contrast thresholds are set based on the quality requirements of the fingerprint image generated by the image generation software.

[0034] It should be noted that the weights of the clarity and contrast are set based on their importance in evaluating the image quality of the reference fingerprint image. In a specific embodiment, the weights of the clarity and contrast are 0.5 and 0.5 respectively.

[0035] S3: Compare the image quality index of each generated reference fingerprint image with a preset image quality index threshold. If the image quality index of a reference fingerprint image is less than the preset image quality index threshold, the quality of the reference fingerprint image does not meet the standard. The reference fingerprint images that do not meet the quality standard are counted and eliminated to obtain the screened reference fingerprint images, which are recorded as high-quality fingerprint images to construct a high-quality fingerprint image set.

[0036] As a preferred solution, the specific analysis process of step 1 further includes: D1: obtaining the fingerprint ridge direction at the pixel point in the filtered high-quality fingerprint image, and recording it as , set the weight of the orthogonal direction of the fingerprint ridge direction at the pixel point in the filtered high-quality fingerprint image, and record it as , set the weight of the fingerprint ridge direction at the pixel point in the filtered high-quality fingerprint image, and record it as .

[0037] In a specific embodiment, the above filtering process is repeated 8-14 times for the same image. and Take 0.9 and 0.1 respectively.

[0038] By analyzing the formula Get the convolution kernel of adaptive filtering ,in , , , and further set the adaptive filtering.

[0039] D2: Use adaptive filtering to destroy the direction field of each high-quality fingerprint image, perform a primary degradation on each high-quality fingerprint, and further perform a secondary degradation on each high-quality fingerprint through Gaussian blur, morphological processing dilation and erosion operations to obtain each fingerprint after the primary and secondary degradation corresponding to each high-quality fingerprint, which are recorded as each degraded fingerprint corresponding to each high-quality fingerprint, and generate a degraded fingerprint set of the high-quality fingerprint set.

[0040] In a specific embodiment, the probability of Gaussian blur is set to 0.5, the blur kernel size is set to 5x5, and the number of repetitions is 8-15 times.

[0041] It should be noted that the fingerprint ridge direction at a pixel point in a high-quality fingerprint image refers to the local direction of the fingerprint ridge line of the high-quality fingerprint image at the image pixel point.

[0042] It should be noted that the adaptive filtering can adjust its convolution kernel according to the fingerprint ridge direction at the pixel point in the high-quality fingerprint image.

[0043] It should be noted that the convolution kernel of the adaptive filtering is the filter parameter of the adaptive filtering.

[0044] As a preferred solution, the specific analysis process of step one also includes: extracting the steel surface defect data set NEU-CLS stored in the database, superimposing the degraded fingerprint set of the high-quality fingerprint set with the steel surface defect data set NEU-CLS to obtain a low-quality latent fingerprint set corresponding to the high-quality fingerprint set, and constructing a paired low-quality latent fingerprint set and a high-quality fingerprint set.

[0045] In a specific embodiment, a dataset with a total of 40,000 pairs of low-quality latent fingerprints and corresponding high-quality fingerprints is constructed.

[0046] In this embodiment, the present invention detects the quality of fingerprint images generated by the software, screens the generated fingerprint images, removes fingerprint images with poor quality, and retains high-quality fingerprint images during the process of degrading high-quality fingerprints generated by the software to obtain paired low-quality latent fingerprints and high-quality fingerprint data sets, thereby improving the quality of the entire fingerprint image database, thereby improving the performance and accuracy of the enhanced network trained based on the high-quality fingerprint set and its derived low-quality latent fingerprint set.

[0047] Step 2: Enhanced network construction: Design a dual-branch spatial attention module and further construct an enhanced network including the dual-branch spatial attention module.

[0048] As a preferred option, see Figure 2 As shown, the specific analysis process of step 2 is: Q1: Use three consecutive residual blocks as high-quality branches of the dual-branch spatial attention module to extract the features of the high-quality area of ​​the input.

[0049] Q2: Use 2 consecutive 2x2 downsampling blocks, 6 residual blocks, and 2 2x2 upsampling blocks as the low-quality branches of the dual-branch spatial attention module to extract features of low-quality regions of the input.

[0050] In a specific embodiment, the downsampling block is composed of 1 3x3 convolution with a stride of 1, padding of 1, and the same number of channels as the input, and 1 3x3 convolution with a stride of 2, padding of 1, and the same number of channels as the input. The upsampling block is composed of 1 3x3 deconvolution with a stride of 2, padding of 1, and the same number of channels as the input, and 1 3x3 convolution with a stride of 1, padding of 1, and the same number of channels as the input. A BatchNorm and ReLU module needs to be added after each convolution of the downsampling block to increase the network's expressive power.

[0051] Q3: Use 3x3 convolution to process the features extracted by the high-quality branch and the low-quality branch of the dual-branch spatial attention module respectively, and add the two processing results to obtain the feature sum.

[0052] In a specific embodiment, the stride of the above 3x3 convolution is 1, the padding is 1, and the number of channels is the number of input channels. .

[0053] Q4: Use 3x3 convolution to process the features and obtained in Q3, and predict the spatial attention weights corresponding to the high-quality branches and low-quality branches.

[0054] In a specific embodiment, the above-mentioned features and are processed using a 3x3 convolution with a stride of 1, a padding of 1, and a channel number of 2 and a Softmax activation function.

[0055] Q5: Multiply and sum the features extracted by the high-quality branch and the low-quality branch with the spatial attention weights corresponding to their branches, and further process the summation results using 6 consecutive residual blocks to obtain the fused features, thereby completing the design of the dual-branch spatial attention module.

[0056] Q6: Add the designed two-branch spatial attention module to the enhanced network to obtain an enhanced network containing the two-branch spatial attention module.

[0057] It should be noted that the spatial attention module is a component used in deep learning to focus on the spatial location information of data such as images and texts. Its main purpose is to enhance the model's perception of features at different spatial locations in the input data, thereby improving the performance of the model.

[0058] It should be noted that the residual block is a basic building block in deep neural networks. It is mainly used to solve the problems of gradient disappearance and gradient explosion in the training process of deep neural networks. At the same time, it enables the network to learn the identity mapping more easily. The core idea is to introduce a "shortcut connection" so that the network can directly pass the input information to the subsequent layers.

[0059] It should be noted that the downsampling block is mainly used to reduce the dimension of the data, usually in the spatial dimension, with the aim of reducing the amount of data while extracting more abstract features.

[0060] It should be noted that the function of the upsampling block is opposite to that of the downsampling block. It is used to increase the dimension of the data, mainly to enlarge it in the spatial dimension, restore the resolution of the data, enable the network to output a result with the same or similar resolution as the input, or to fuse features of different scales.

[0061] It should be noted that 3x3 convolution is a convolution operation in convolutional neural networks, which uses a convolution kernel, i.e., a filter, to slide on input data such as images, feature maps, etc. to calculate the convolution result.

[0062] It should be noted that the designed dual-branch spatial attention module is used to simultaneously learn the features of high-quality fingerprint regions and low-quality fingerprint regions.

[0063] It should be noted that the enhanced network containing the dual-branch spatial attention module can simultaneously extract features of different quality areas, thereby improving the expressiveness of the enhanced network and the final enhanced performance.

[0064] In this embodiment, the present invention designs a dual-branch spatial attention module so that the spatial attention can fuse the different features extracted by the high-quality branch and the low-quality branch, and further adds the dual-branch spatial attention module to the enhanced network, so that the enhanced network can simultaneously pay attention to fingerprint areas of different qualities and learn the features of high-quality and low-quality fingerprint areas, thereby improving the enhancement effect of the enhanced network on latent fingerprints.

[0065] Step 3. Generate the total loss of the enhanced network: Design the reconstruction loss and each directional field loss of the enhanced network, where each directional field loss includes the directional field loss of the prediction result of the high-quality branch, the directional field loss of the prediction result of the low-quality branch, and the directional field loss of the prediction result after the corresponding features of the high-quality branch and the low-quality branch are fused, and further generate the total loss of the enhanced network.

[0066] As a preferred option, see Figure 3 As shown, the specific analysis process of step three includes: by analyzing the formula Get the reconstruction loss of the enhanced network ,in Indicates the amount of data used for each weight update in each round of training, It represents the result of enhancing the low-quality latent fingerprint input using the enhancement network. represents the high-quality fingerprint corresponding to the input low-quality latent fingerprint, represents the L1 norm.

[0067] It should be noted that the use of The loss design enhances the reconstruction loss of the network, which can improve the ridge-valley contrast of the enhanced fingerprint.

[0068] As a preferred solution, the specific analysis process of step three also includes: T1: adding a 3x3 convolution after the high-quality branch, low-quality branch and two-branch feature fusion stage of the dual-branch spatial attention module of the enhanced network, respectively, to obtain the direction field predicted after the high-quality branch, low-quality branch and two-branch feature fusion, and record them as .

[0069] In a specific embodiment, the stride of the 3x3 convolution is 1, the padding is 1, and the number of channels is 2.

[0070] T2: By analyzing the formula Direction field loss for high-quality branch prediction results ,in represents the direction field of the high-quality fingerprint corresponding to the input low-quality latent fingerprint, represents the L2 norm.

[0071] T3: By analyzing the formula Direction field loss for low quality branch prediction results .

[0072] T4: By analyzing the formula The directional field loss of the prediction result after the fusion of the corresponding features of the high-quality branch and the low-quality branch is obtained .

[0073] As a preferred solution, the specific analysis process of step 3 also includes: by analyzing the formula The total loss of the enhanced network is obtained ,in , and They respectively represent the weights of the preset high-quality branch, low-quality branch, and the direction field corresponding to the fusion features of the high-quality branch and the low-quality branch.

[0074] In a specific embodiment, , and Take 0.5, 0.8 and 1.0 respectively.

[0075] In this embodiment, the present invention guides the training of the enhancement network by using reconstruction loss and multiple directional field losses, so that the enhancement network can learn the directional field characteristics of the latent fingerprint, improves the learning ability of the enhancement network for the ridge directional characteristics of the latent fingerprint, and thus improves the accuracy of the enhancement network in identifying latent fingerprints.

[0076] Step 4: Enhanced network training: Use the paired low-quality latent fingerprint set and high-quality fingerprint set and the total loss of the enhanced network to train the enhanced network to obtain the trained enhanced network.

[0077] As a preferred solution, the specific analysis process of step 4 is as follows: A1: dividing the paired low-quality latent fingerprint set and high-quality fingerprint set to obtain a training data set and a verification data set for the enhanced network.

[0078] In a specific embodiment, the constructed 40,000 pairs of low-quality latent fingerprint sets and corresponding high-quality fingerprint sets are divided into a training data set with 35,000 pairs of data and a validation data set with 5,000 pairs of data.

[0079] A2: Set the operating parameters of the enhanced network training process, where the operating parameters include GPU memory size, optimizer type, initial learning rate, number of data pairs used for each parameter update, and total number of iterations. Further, train the enhanced network based on the training data set of the enhanced network, and record the enhanced network obtained through training as the quasi-enhanced network.

[0080] In a specific embodiment, during the enhanced network training process, a GPU with a video memory of 24 GB is selected for training, the optimizer is the Adam optimizer, and the initial learning rate is , each parameter update uses 8 pairs of data, with a total of 43750 iterations.

[0081] A3: According to the validation data set and total loss of the enhanced network, the total loss of the proposed enhanced network on the validation data set is calculated to determine whether the total loss of the proposed enhanced network on the validation data set is the smallest. If it is the smallest, the proposed enhanced network is recorded as the final enhanced network. Otherwise, the training process of step A2 is repeated until the training result with the smallest total loss on the validation data set is obtained, which is used as the final enhanced network, thereby obtaining the trained enhanced network.

[0082] It should be noted that whether the total loss of the proposed enhanced network on the validation data set is the smallest is determined by comparing the total losses of the proposed enhanced network on the validation data set obtained by executing the training process each time.

[0083] It should be noted that during the training of the enhanced network, the training result with the smallest total loss on the validation data set will be used as the final enhanced network.

[0084] Step 5: Latent fingerprint image enhancement: Input the latent fingerprint image to be enhanced into the trained enhancement network, obtain the enhanced latent fingerprint image, and provide feedback.

[0085] As a preferred option, see Figure 4 As shown, the specific analysis process of step 5 is as follows: B1: Use three consecutive convolution-downsampling blocks to preliminarily extract the features of the latent fingerprint image to be enhanced, the number of image feature channels becomes 256, and the image size becomes the input .

[0086] In a specific embodiment, the first convolution-downsampling block is composed of 1 3x3 convolution with a stride of 1, padding of 1, and 64 channels, and 1 3x3 convolution with a stride of 2, padding of 1, and 64 channels; the second convolution-downsampling block is composed of 1 3x3 convolution with a stride of 1, padding of 1, and 128 channels, and 1 3x3 convolution with a stride of 2, padding of 1, and 128 channels; the third convolution-downsampling block is composed of 1 3x3 convolution with a stride of 1, padding of 1, and 256 channels, and 1 3x3 convolution with a stride of 2, padding of 1, and 256 channels. A BatchNorm and ReLU module need to be added after each convolution of the convolution-downsampling block to increase the expression ability of the network.

[0087] B2: Input the features of the latent fingerprint image to be enhanced extracted in step B1 into the dual-branch spatial attention module, extract the features of different quality areas of the latent fingerprint image to be enhanced and fuse them.

[0088] B3: Use three consecutive deconvolution-upsampling blocks to process the result of step B2 to obtain an enhancement result with the same number of channels and the same size as the input latent fingerprint to be enhanced, and then obtain the enhanced latent fingerprint image for feedback.

[0089] In a specific embodiment, the first deconvolution-upsampling block is composed of 1 3x3 deconvolution with a stride of 2, padding of 1, and 256 channels, and 1 3x3 convolution with a stride of 1, padding of 1, and 256 channels; the second deconvolution-upsampling block is composed of 1 3x3 deconvolution with a stride of 2, padding of 1, and 128 channels, and 1 3x3 convolution with a stride of 1, padding of 1, and 128 channels; the third deconvolution-upsampling block is composed of 1 3x3 deconvolution with a stride of 2, padding of 1, and 64 channels, and 1 3x3 convolution with a stride of 1, padding of 1, and 1 channel. A BatchNorm and ReLU module need to be added after each convolution of the deconvolution-upsampling block to increase the expression ability of the network.

[0090] The above contents are merely examples and explanations of the concept of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the present invention, they shall all fall within the protection scope of the present invention.

Claims

1. A fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology, characterized in that: The steps include: Step 1, obtaining paired low-quality latent fingerprint sets and high-quality fingerprint sets: obtaining a high-quality fingerprint image set, degrading each high-quality fingerprint once through adaptive filtering, further degrading each high-quality fingerprint twice through Gaussian blur, morphological processing expansion and corrosion operations, generating a degraded fingerprint set of the high-quality fingerprint set, superimposing the degraded fingerprint set of the high-quality fingerprint set with the steel surface defect data set, obtaining a low-quality latent fingerprint set corresponding to the high-quality fingerprint set, and constructing paired low-quality latent fingerprint set and high-quality fingerprint set; Step 2: Enhanced network construction: Design a dual-branch spatial attention module and further construct an enhanced network containing the dual-branch spatial attention module; The specific analysis process of step 2 is as follows: Q1: Use three consecutive residual blocks as high-quality branches of the dual-branch spatial attention module to extract features of high-quality regions of the input; Q2: Use two consecutive 2x2 downsampling blocks, six residual blocks, and two 2x2 upsampling blocks as the low-quality branches of the dual-branch spatial attention module to extract features of low-quality regions of the input; Q3: Use 3x3 convolution to process the features extracted by the high-quality branch and the low-quality branch of the dual-branch spatial attention module respectively, and add the two processing results to obtain the feature sum; Q4: Use 3x3 convolution to process the features and obtained in Q3, and predict the spatial attention weights corresponding to the high-quality branch and the low-quality branch; Q5: Multiply and sum the features extracted by the high-quality branch and the low-quality branch with the spatial attention weights corresponding to their branches, and further process the summation results using 6 consecutive residual blocks to obtain the fused features, thereby completing the design of the dual-branch spatial attention module; Q6: Add the designed dual-branch spatial attention module to the enhanced network to obtain an enhanced network containing the dual-branch spatial attention module; Step 3: Generate the total loss of the enhanced network: Design the reconstruction loss and each directional field loss of the enhanced network, where each directional field loss includes the directional field loss of the high-quality branch prediction result, the directional field loss of the low-quality branch prediction result, and the directional field loss of the prediction result after the corresponding features of the high-quality branch and the low-quality branch are fused, and further generate the total loss of the enhanced network; Step 4: Enhanced network training: Use the paired low-quality latent fingerprint set and high-quality fingerprint set and the total loss of the enhanced network to train the enhanced network and obtain the trained enhanced network; Step 5: Latent fingerprint image enhancement: Input the latent fingerprint image to be enhanced into the trained enhancement network, obtain the enhanced latent fingerprint image, and provide feedback.

2. The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology according to claim 1 is characterized in that: The specific analysis process of step 1 includes: S1: Setting technical parameter information for generating high-quality fingerprint images, wherein the technical parameter information includes prompt word information, generation parameter information, and model type, and inputting them into Anguli image generation software to generate a set number of reference fingerprint images; S2: Detect each generated reference fingerprint image to obtain the clarity and contrast of each reference fingerprint image, which are recorded as , Indicates The number of the reference fingerprint image, , by analyzing the formula Get the image quality index of each generated reference fingerprint image ,in represents a natural constant, represents the correction factor of the preset image quality index, Respectively represent the preset clarity and contrast thresholds, Respectively represent the preset clarity and contrast weights, ; S3: Compare the image quality index of each generated reference fingerprint image with a preset image quality index threshold. If the image quality index of a reference fingerprint image is less than the preset image quality index threshold, the quality of the reference fingerprint image does not meet the standard. The reference fingerprint images that do not meet the quality standard are counted and eliminated to obtain the screened reference fingerprint images, which are recorded as high-quality fingerprint images to construct a high-quality fingerprint image set.

3. The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology according to claim 2 is characterized in that: The specific analysis process of step 1 also includes: D1: Get the fingerprint ridge direction at the pixel point in the high-quality fingerprint image to be filtered, and record it as , set the weight of the orthogonal direction of the fingerprint ridge direction at the pixel point in the filtered high-quality fingerprint image, and record it as , set the weight of the fingerprint ridge direction at the pixel point in the filtered high-quality fingerprint image, and record it as ; By analyzing the formula Get the convolution kernel of adaptive filtering ,in , , , further set the adaptive filtering; D2: Use adaptive filtering to destroy the direction field of each high-quality fingerprint image, perform a primary degradation on each high-quality fingerprint, and further perform a secondary degradation on each high-quality fingerprint through Gaussian blur, morphological processing dilation and erosion operations to obtain each fingerprint after the primary and secondary degradation corresponding to each high-quality fingerprint, which are recorded as each degraded fingerprint corresponding to each high-quality fingerprint, and generate a degraded fingerprint set of the high-quality fingerprint set.

4. The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology according to claim 3 is characterized in that: The specific analysis process of step 1 also includes: The steel surface defect dataset NEU-CLS stored in the database is extracted, and the degraded fingerprint set of the high-quality fingerprint set is superimposed with the steel surface defect dataset NEU-CLS to obtain the low-quality latent fingerprint set corresponding to the high-quality fingerprint set, and the paired low-quality latent fingerprint set and high-quality fingerprint set are constructed.

5. The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology according to claim 1 is characterized in that: The specific analysis process of step three includes: By analyzing the formula Get the reconstruction loss of the enhanced network ,in Indicates the amount of data used for each weight update in each round of training, It represents the result of enhancing the low-quality latent fingerprint input using the enhancement network. represents the high-quality fingerprint corresponding to the input low-quality latent fingerprint, represents the L1 norm.

6. The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology according to claim 5 is characterized in that: The specific analysis process of step three also includes: T1: A 3x3 convolution is added after the high-quality branch, low-quality branch, and two-branch feature fusion stage of the dual-branch spatial attention module of the enhanced network to obtain the direction field predicted after the high-quality branch, low-quality branch, and two-branch feature fusion, which are recorded as ; T2: By analyzing the formula Direction field loss for high-quality branch prediction results ,in represents the direction field of the high-quality fingerprint corresponding to the input low-quality latent fingerprint, represents the L2 norm; T3: By analyzing the formula Direction field loss for low quality branch prediction results ; T4: By analyzing the formula The directional field loss of the prediction result after the fusion of the corresponding features of the high-quality branch and the low-quality branch is obtained .

7. The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology according to claim 6 is characterized in that: The specific analysis process of step three also includes: By analyzing the formula The total loss of the enhanced network is obtained ,in , and They respectively represent the weights of the preset high-quality branch, low-quality branch, and the direction field corresponding to the fusion features of the high-quality branch and the low-quality branch.

8. The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology according to claim 1 is characterized in that: The specific analysis process of step 4 is as follows: A1: Divide the paired low-quality latent fingerprint set and high-quality fingerprint set to obtain the training data set and verification data set of the enhanced network; A2: Set the operating parameters of the enhanced network training process, where the operating parameters include GPU memory size, optimizer type, initial learning rate, number of data pairs used for each parameter update, and total number of iterations. Further, train the enhanced network based on the training data set of the enhanced network, and record the enhanced network obtained through training as the quasi-enhanced network. A3: According to the validation data set and total loss of the enhanced network, the total loss of the proposed enhanced network on the validation data set is calculated to determine whether the total loss of the proposed enhanced network on the validation data set is the smallest. If it is the smallest, the proposed enhanced network is recorded as the final enhanced network. Otherwise, the training process of step A2 is repeated until the training result with the smallest total loss on the validation data set is obtained, which is used as the final enhanced network, thereby obtaining the trained enhanced network.

9. The fingerprint feature recognition and analysis method based on directional field guidance and spatial attention technology according to claim 1 is characterized in that: The specific analysis process of step 5 is as follows: B1: Use three consecutive convolution-downsampling blocks to preliminarily extract the features of the latent fingerprint image to be enhanced. The number of image feature channels becomes 256, and the image size becomes the input ; B2: Input the features of the latent fingerprint image to be enhanced extracted in step B1 into the dual-branch spatial attention module, extract the features of different quality areas of the latent fingerprint image to be enhanced and fuse them; B3: Use three consecutive deconvolution-upsampling blocks to process the result of step B2 to obtain an enhancement result with the same number of channels and the same size as the input latent fingerprint to be enhanced, and then obtain the enhanced latent fingerprint image for feedback.

Images